80 results about "Neural network architecture" patented technology

The application provides a lightweight neural network architecture search method based on flatness, which is used for reducing the time cost and computing resources of searching the neural network architecture. From the perspective of predicting the network generalization, the application provides a method for comparing the networks by taking the flatness of the candidate network at the initial time as an evaluation index. The method comprises the following steps: determining an image processing task and determining a data set; determining a network search space; selecting a certain number of neural network architectures in the search space; verifying the effectiveness of the evaluation index; calculating the flatness of the neural network architecture, and sorting the selected architectures according to the flatness; selecting the architecture with the maximum flatness as the optimal architecture; and training the optimal architecture to complete the image classification task. The application can select an architecture with high accuracy in a small time consumption.

The embodiment of the application relates to the technical field of dance generation, and provides a dance generation model training method and a dance generation method, a deep-coupled neural network architecture is constructed, a dynamics constraint module based on a space-time graph neural network is innovatively introduced in a latent space of a variational autoencoder, specific dynamics bias is explicitly predicted and applied to an action intention according to a specified clothing type, dance action sequences conforming to physical laws and having specific clothing dynamics are forced to be generated by the network output, the generated dance action sequences can not only match music rhythm, but also can truly present physical dynamics of the specified clothing, and the authenticity and expressiveness of dance generation under complex clothing are significantly improved.

This invention discloses a method for restoring turbulent degradation images based on complex amplitude detection, belonging to the field of computational optics imaging technology. This method constructs a computational optics imaging system without a hardware correction unit. It utilizes a wavefront sensor and an imaging sensor to simultaneously acquire Hartmann images and target degradation images. A computational unit performs wavefront complex amplitude reconstruction and aberration correction. Based on this, a physically constrained end-to-end neural network architecture is designed. By reconstructing the wavefront complex amplitude, it achieves accurate mapping and inference from the Hartmann image to the point spread function of the imaging system. Finally, a multi-scale deconvolutional network is combined to complete high-quality restoration of the turbulent degradation image. This invention eliminates the need for a hardware corrector, overcomes the physical limitations of traditional adaptive optics, effectively solves the problem of insufficient point spread function accuracy, and enables time-delay-free and highly efficient turbulent image restoration under low-cost conditions.

Embodiments of the present application provide a training method and device of an image processing model, and a computer readable storage medium, the image processing model comprising a U-shaped network, the U-shaped network comprising a down-sampling path and an up-sampling path, the down-sampling path comprising a first down-sampling module, the up-sampling path comprising a first up-sampling module, the first up-sampling module being connected to the first down-sampling module to multiplex features output by the first down-sampling module, the method comprising: inputting training data into the image processing model to obtain a training loss of the image processing model; updating model parameters of the image processing model according to the training loss; in the process of updating the model parameters, performing neural network architecture search on a network architecture between the first down-sampling module and the first up-sampling module to determine a structure of an attention module between the first down-sampling module and the first up-sampling module.

The application discloses a pathological image segmentation method based on a two-stage differentiable super-labeled attention neural network architecture search, and comprises the following steps: constructing a search space of a two-stage differentiable super-labeled attention module according to pathological image features, wherein the search space comprises a first-stage search space and a second-stage search space; adopting a double-layer optimization strategy to jointly optimize structure parameters and network weights according to the search space, wherein network weights and multi-head adaptive parameters are updated based on a training set loss, and architecture parameters are updated based on a validation set loss; discretizing the first-stage search space into a determined self-attention mechanism branch according to the optimized architecture parameters, fixing a multi-head weighted structure in the second-stage search space, and obtaining an optimized two-stage differentiable super-labeled attention module; and constructing a pathological image segmentation model according to the optimized two-stage differentiable super-labeled attention module, segmenting pathological images, and obtaining a segmentation result. The application can improve segmentation precision and generalization performance.

Methods, systems, and devices for wireless communication are described. A user equipment (UE) can receive an indication of one or more encoding operations for encoding a compressed data set, the one or more encoding operations including a differential encoding operation or an entropy encoding operation or both. In some examples, using a neural network, the UE can first encode the data set based on an additional encoding operation to generate a compressed data set, and then quantize the compressed data set encoded based on the additional encoding operation. Subsequently, after the data set has been initially encoded and then quantized, the UE can further encode and compress the data set using the indication of the one or more encoding operations. The UE can then transmit the data set to a second device based on the one or more encoding operations.

The application proposes a method for realizing atrial fibrillation prediction by using a neural network with a multi-scale attention mechanism, comprising the following steps: step S1: collecting, dividing and pre-processing atrial fibrillation electrocardio signal (AFECG) data set and normal sinus rhythm electrocardio signal (NSR ECG) data set; step S2: designing a neural network architecture to preliminarily predict atrial fibrillation data, and optimizing the prediction network structure on the basis of the network architecture; step S3: constructing a lead attention mechanism among different leads of atrial fibrillation prediction data; step S4: constructing a time-space attention mechanism among different feature maps in the neural network; step S5: constructing a time sequence attention mechanism on different time sequence segments of atrial fibrillation prediction data; step S6: after the construction of each attention mechanism sub-class module is completed, the sub-class module is fused with the basic neural network, and then the overall optimization of the neural network is carried out to form a final atrial fibrillation prediction network; the application can improve the prediction accuracy of atrial fibrillation.

This invention belongs to the field of digital holographic imaging and artificial intelligence technology, specifically a holographic particle recognition method based on EFPFNet. While maintaining the traditional holographic optical path structure, this invention addresses the problems of low target contrast, blurred edges, and strong background interference in micro-nano particle holograms. Unlike the traditional approach of directly transferring general target detection models, it proposes a neural network architecture specifically for holographic image features: employing an edge feature convolution (EFC) module and a particle focusing mechanism (PFM) module to enhance the perception of weak particle diffraction features and suppress complex background noise, respectively; and using simulation data generated based on a physical model as the training set, effectively solving the problem of scarce real labeled data. This method possesses high detection accuracy, high robustness, and excellent generalization ability, and can be applied to target detection scenarios such as biomedicine and industrial inspection without significantly increasing optical hardware costs or detection time.

Method and apparatus for processing data using a reconfigurable, hyperdimensional neural network architecture comprising a feature extractor and a classifier. The feature extractor comprises a neural network for encoding input information into hyperdimensional (HD) vectors and extracting at least one particular HD vector representing at least one feature within the input information, wherein the neural network comprises no more than one multiply and accumulate operator. The classifier is coupled to the feature extractor for classifying the at least one particular HD vector to produce an indicium of classification for the at least one particular HD vector and wherein the classifier does not comprise any multiply and accumulate operators.

This invention belongs to the field of 5G communication technology, specifically a wireless resource allocation method based on the deep reinforcement learning (DQN) algorithm under the 5G standard. The invention includes: combining wireless resources with dual-layer coding technology to model the panoramic video experience quality of individual users; fully considering user experience quality requirements and the heterogeneity of user channel states to determine the order of user resource allocation; modeling state information and user information; designing a neural network architecture and combining it with the deep reinforcement learning (DQN) algorithm to allocate appropriate optional parameter sets and minimum time slots to each user, thereby maximizing the overall panoramic video experience quality while meeting the basic experience quality requirements of all users. This invention can provide higher scalability and practicality for wireless resource allocation at the physical layer level, improve the utilization rate of limited communication resources, and has broad application prospects.

This application provides an image processing method and apparatus based on adversarial neural network (DNN) architecture search. The method includes: for any epoch in the DNN architecture search process, iteratively updating the operating parameters and structural parameters of the DNN network using a gradient descent algorithm, an acquired image training set, and an acquired image validation set until the number of iterations reaches a first iteration number; and iteratively updating the structural parameters of the DNN network using preset network vulnerability constraints and the acquired image validation set until the number of iterations within that epoch reaches a second iteration number; when the number of searched epochs reaches the first epoch number, or when the DNN network model converges, generating a target DNN network for image processing based on the obtained structural parameters, and using the target DNN network to process the image to be processed. This method can improve the accuracy of image processing using DNN networks.

The present application belongs to the field of soil seepage and erosion model parameter calibration, and discloses a soil seepage and erosion model parameter calibration method and system based on multi-fidelity data neural network and CFD-DEM coupling simulation, comprising: obtaining low-fidelity data and high-fidelity data, and performing data normalization processing; building a neural network architecture, and calculating macroscopic seepage and erosion parameters through deep learning inversion; and performing finite element simulation integration. The present application calibrates the seepage and erosion law parameters of soil through the generated low-fidelity data combined with a small amount of high-fidelity data obtained through on-site experiments. The Softplus / ReLU activation function is introduced to force the parameters to be non-negative, eliminating the risk of non-physical results of model output. The hidden layers of the high-fidelity neural network are divided into linear hidden layers and nonlinear hidden layers, and the complex relationship between low-fidelity and high-fidelity data is adaptively adjusted and learned. The present application is easy to use and responds quickly, and after training and deployment, employees with short-term training can operate without professional personnel.

The application discloses a neural network-based operator home broadband complaint prediction method and system, and relates to the technical field of big data mining and analysis. Since home broadband problems are difficult to be quickly solved, the scheme comprises the following steps: obtaining user complaint work orders, extracting device state information and alarm information at the time of user complaints, constructing feature engineering and generating a supervised learning time series dataset after mining and preprocessing; dividing the dataset, designing a neural network architecture based on LSTM, training the neural network by using the divided dataset, and finally outputting an LSTM multivariate time series prediction model; optimizing the training results of the prediction model, introducing a time series attention mechanism and a combined loss function, and improving the performance of the model in adapting to the complaint prediction task; and deploying the optimized prediction model to an actual business system, positioning potential complaint customers and predicting complaint reasons by accessing real-time device state information and alarm information. The application can improve the efficiency of on-site handling by installation and maintenance personnel.

A system includes a neural network architecture. It a new type of neural network able to process statically mapped as well as temporally sequenced information with much better power utilization, data requirements and operational efficiencies. Unlike prior artificial neural network approaches, the present invention includes uniquely defined sets of relationships. The unique use of non-linear input-output mapping functions combined with a time-variant pilot function, and a deterministically bounded, fully-differentiable, nonlinear resonance field subsystem allows the present invention to be readily deployed to work with virtually any neural network architecture / implementation including photonic, opto-acoustic or other variants. This dramatically reduces the size and complexity of virtually any neural network architecture because it offloads what would otherwise need to be done in the form of back / forward propagation trained weights and biases to much simpler, more scalable differentiable input / output mapping functions.

This invention relates to an intelligent detection method for the dynamic process of water accumulation on highway pavements, belonging to the fields of computer vision and intelligent transportation. The method includes: acquiring videos from multiple sections of the highway; preprocessing the video frames of each video and calculating the dense optical flow of adjacent frames to obtain an optical flow sequence; constructing a dual-flow fusion neural network architecture; integrating a watershed attention mechanism into the spatial flow network to extract spatial features; using a 3D convolutional network to process the optical flow sequence to extract temporal features; adaptively calculating the fusion weights of the spatial and temporal flows to generate fusion features; using a multi-task learning framework to generate corresponding prediction results; using a physical perception loss function to constrain the prediction results; and performing data analysis and risk assessment on the constrained prediction results. This invention's method can detect the dynamic process of water accumulation on highway pavements in real time and accurately, providing important technical support for traffic safety.

The application discloses a seawater aging microplastic identification method based on a multimodal deep learning, relates to the technical field of environmental monitoring and artificial intelligence, and comprises the following steps: preparing a plurality of seawater microplastic samples, performing an aging experiment on the microplastic samples, regularly collecting spectrum data and microscopic images of the samples during the aging process, and constructing a data set; performing noise reduction, baseline correction, normalization and dimension reduction processing on the collected spectrum data, and processing the microscopic images by using an image processing algorithm; constructing a double-branch deep learning model, adopting a light-weight improved convolutional neural network architecture for a visual branch, adopting a sparse principal component analysis dimension reduction combined with a full-connection neural network architecture for a spectrum branch, fusing the features of the two branches by using a feature fusion module, and outputting an identification result by using a classifier; and dividing the data set into a training set, a verification set and a test set, training the model, and obtaining a final model. The application has high identification precision, strong robustness, high automation and strong practicability.

Method for generating an optimized neural network architecture taking into account the hardware constraints of a hardware target. The present invention relates to a method for generating an optimized neural network architecture taking into account the hardware constraints of a hardware target (20), the method comprising: obtaining an algorithmic task (100); defining a neural search space (104) as a function of the algorithmic task (100); defining a hardware configuration space (102) of the hardware target; searching for a neural network architecture for the hardware target using a genetic algorithm (108) to optimize evaluation criteria (106), the evaluation criteria (106) being at least one hardware performance and at least one algorithmic performance;and evaluate the hardware performance(s) by using a performance predictor if the performance predictor has a reliability above a reliability threshold or the hardware performance(s) being determined by a performance evaluation by running the neural network on the hardware target (20) if the performance predictor has a reliability below the reliability threshold. Figure for the abstract: 2;

Methods and systems are described for real-time coaching and guidance using a virtual assistant that interacts with a user. A user can receive feedback inferences provided substantially in real-time after a video sample is collected from the user's device. Neural network architectures and layers can be used to determine motion patterns and temporal aspects of the video sample, as well as to detect the activity of the foreground user despite background noise. The methods and systems can have various capabilities, including but not limited to real-time feedback on the exercise activity performed, exercise score, calorie estimation, and repetition count.

Exemplary system and methods use a combination of application modules and neural network architecture for multi-speaker and multi-language speech analysis. The exemplary system can receive a natural language input, which it decomposes into plural segments. A sub-group of the plural segments are accumulated in a buffer where each segment representing a period during which voice activity is detected. The sub-groups are analyzed for voice activity of multiple speakers and one or more text segments are generated based on the speakers. A semantic vector for each text segment is generated and stored in vector memory. Relevant data associated with each semantic vector is retrieved from the vector memory based on a similarity measure; and a response including specified information extracted from the one or more text segments is generated based on at least the relevant data.

This invention proposes a method and system for constructing an intelligent agent for reviewing reports on a large-scale model of port facility maintenance. The method includes: collecting cross-modal raw datasets; constructing a multimodal feature fusion perception layer to generate facility damage feature aligned data; constructing a four-level cognitive neural network architecture to generate an inference decision tree; constructing a root cause-path-effect causal chain to generate a fault attribution analysis report; constructing a prediction-intervention-verification proactive defense closed loop to generate a Pareto optimal maintenance strategy set; and executing the intervention strategy and providing feedback on the verification results using a digital twin verification platform and a blockchain evidence storage system. This invention achieves deep semantic alignment of cross-modal data through a multimodal feature fusion perception layer, breaking down data silos and improving the accuracy of damage feature extraction; constructs an interpretable causal chain based on relevant architectures and modules to solve the black box problem of decision-making, providing causal logic support for maintenance strategies; and achieves real-time verification and reliable traceability of strategy effects through a proactive defense closed loop.

This invention relates to the fields of automated machine learning and neural network technology, specifically to a method and system for searching neural network structures based on dynamic collaborative graph encoding. This method leverages the collaborative evolution mechanism of maintainer, sterile, and restorer lines in breeding optimization algorithms, achieving inter-population structural migration through dynamic hybridization probability control. A support vector machine surrogate model is constructed to replace time-consuming performance evaluation, rapidly predicting the accuracy of offspring structures and screening high-potential candidate structures. Gradient fine-tuning is applied to candidate structures to output the optimal network structure. Modular recombination and topological evolution of the neural network structure are achieved through graph encoding. Utilizing the collaborative mechanism of maintaining the stability of the maintainer line, exploring the diversity of the sterile line, and accelerating the convergence of the restorer line, combined with a two-layer dynamic regulation of population-level migration control and individual-level variation optimization, this method efficiently obtains high-performance deep neural network structures.

The application discloses a wind power prediction method and system fusing wind field images and multiple physical constraints, and belongs to the technical field of new energy power generation, aiming to improve the accuracy and stability of short-term wind power prediction. The method comprises the following steps: obtaining wind field image data and performing pretreatment; extracting features by using a dynamic time sequence encoder, wherein the dynamic time sequence encoder comprises two parallel branches of a Fourier neural operator and a convolution gate recurrent unit, which are respectively used for capturing global frequency spectrum features and local transient features of the wind field; introducing spatial variable gate parameters and learnable channel weights to fuse features of each branch; based on the fused features, a neural network architecture driven by physical information is used to realize multi-step wind field simulation; and multi-step wind power prediction is performed on multiple wind farms by using a time sequence fusion Transformer according to multi-step wind field images obtained through the wind field simulation. The application overcomes the shortcomings of existing methods in capturing the space-time dependence and local rapid disturbance of the wind field.

This invention discloses an intelligent optimization method and system for industrial energy consumption based on multi-source data fusion, relating to the field of intelligent manufacturing technology. The method includes: collecting multi-source heterogeneous operational data from industrial sites, generating a standardized fusion dataset, and initializing neural network architecture parameters; iteratively training the neural network architecture parameters to be optimized using the standardized fusion dataset, quantifying the degree of deviation between data prediction error terms and physical equation residual terms in real time during training to construct a mechanism-data conflict index; dynamically relaxing the weight coefficients of physical constraint terms based on the conflict index to balance the game relationship between measured data trends and theoretical physical constraints, thereby obtaining an optimized neural network model; and outputting a virtual state vector using the optimized neural network model; driving a digital twin instance to run, injecting the virtual state vector in real time to maintain virtual-real mapping synchronization, and forming a digital twin simulation environment.

The application relates to the field of geomagnetic vector measurement, and discloses a magnetic target intelligent positioning method and system based on a layered neural network architecture. The method comprises the following steps: acquiring multi-point magnetic field data measured by a magnetic sensor array; calculating magnetic gradient tensor data according to the magnetic field data and the spatial position of the sensor; inputting the magnetic gradient tensor data into a pre-constructed layered neural network model; extracting local space-time features from the magnetic gradient tensor time series data of each measurement point through a sub-network layer; performing time series fusion and global context modeling on the feature sequences output by all sub-network branches through a global fusion network layer; and outputting the three-dimensional spatial position coordinates of the magnetic target according to the fused global features through an output layer. The implementation of the application is a deep integration of physical analysis methods and data-driven methods, can guide the mapping relationship between the magnetic gradient tensor and the position of the magnetic target of the deep learning model, and retains the strong nonlinear fitting capability, thereby enhancing the positioning accuracy of the magnetic target.

This invention relates to the field of neural network architecture search and industrial intelligent processing technology, and discloses a structure search method for multi-output dendritic neuron models for industrial classification tasks. The method includes: initializing the synaptic connection weights and dendritic threshold parameters of the multi-input multi-output dendritic neuron model to generate an initial population; uniformly dividing the initial population into several subpopulations of equal size; calculating the temporal and spatial criteria for each subpopulation after task allocation and optimization, and dynamically allocating evolutionary computational resources for each subpopulation based on the fitness change rate represented by the temporal criterion and the population distribution state represented by the spatial criterion; implementing an accelerated sharing penalty mechanism to adjust fitness values ​​according to the crowding degree among individuals to maintain solution set diversity; and updating each subpopulation. This invention solves the problem of low convergence efficiency caused by uniform resource allocation in traditional large-scale multi-objective evolutionary algorithms during neural network architecture search.

This invention discloses a machine learning optimization method based on a mathematical model, relating to the field of machine learning technology. The method is implemented through the following steps: First, optimization points and momentum are initialized on a Lie group; then, gradient calculation, momentum update, and point update operations are repeatedly performed until the convergence condition is met; subsequently, the iterative operations are mapped to optimization layers in a neural network, and multiple optimization layers are stacked to construct a deep unfolded neural network architecture; this architecture is trained using domain-specific data to optimize network parameters; finally, the trained network is deployed to optimize mathematical models with geometrically structured data in robotics. This invention fully utilizes the geometrical characteristics of the data, improving the efficiency, accuracy, and stability of the optimization process, enhancing the model's understanding and adaptability to problems, and can be applied to multiple problems in robotics, such as posture estimation and trajectory tracking, to achieve effective optimization.

The present disclosure provides directed to new, more efficient neural network architectures. As one example, in some implementations, the neural network architectures of the present disclosure can include a linear bottleneck layer positioned structurally prior to and / or after one or more convolutional layers, such as, for example, one or more depthwise separable convolutional layers. As another example, in some implementations, the neural network architectures of the present disclosure can include one or more inverted residual blocks where the input and output of the inverted residual block are thin bottleneck layers, while an intermediate layer is an expanded representation. For example, the expanded representation can include one or more convolutional layers, such as, for example, one or more depthwise separable convolutional layers. A residual shortcut connection can exist between the thin bottleneck layers that play a role of an input and output of the inverted residual block.

The application belongs to the technical field of integrated circuits, and particularly relates to a single-channel analog-to-digital converter calibration system based on a memory-aware gray-box neural network. The calibration system adopts a double-layer dynamic gray-box neural network architecture, which comprises: an input layer, which is used for receiving digital output codes of an analog-to-digital converter and performing bit weight-related scaling processing; a recursive storage layer, which is used for storing output information of the analog-to-digital converter at a previous moment; a hidden processing module, which adopts a rectified linear unit as an activation function to perform nonlinear transformation on the input information; and an output layer module, which adopts a multipath data selector to replace a multiplier to generate a calibrated digital output signal. The application can realize efficient collaborative compensation of nonlinear characteristics and memory effects of the analog-to-digital converter, and significantly reduce the calculation complexity and power consumption of the calibration circuit. While ensuring calibration accuracy, the application has the characteristics of fast convergence speed and low hardware overhead.

This invention discloses a global context-aware sequence data processing method and neural network architecture, comprising the following steps: Step 1) acquiring the input sequence; Step 2) using an embedding module to transform the input sequence block into a dense word vector representation; Step 3) using a positional encoding module to inject the absolute or relative position information of each word vector in the input sequence, and generating a processed sequence vector; Step 4) based on the processed sequence vector, an encoder stack generates a high-level sequence representation carrying global context information; Step 5) the decoder stack, based on the output of the encoder stack and the generated partial output sequence, progressively generates the target sequence in an autoregressive manner. This invention, through a global context-aware (self-attention) mechanism, allows all elements in the sequence to perform matrix operations simultaneously when calculating relevance, achieving true data-parallel intra-layer computation.

The application belongs to the technical field of image processing, and specifically discloses a two-stage full 3D aneurysm segmentation method and system. The method acquires a 3D aneurysm detection image and performs pretreatment. A random data enhancement method is used to perform data enhancement on the pretreated 3D aneurysm detection image. A plurality of samples are generated on the data-enhanced 3D aneurysm detection image. A coarse detection network is optimized through neural network architecture search. Based on the plurality of samples, the coarse detection network is trained, and a region of interest in the sample is extracted. The region of interest is taken as input to train a fine segmentation network. The fine segmentation network segments an aneurysm in the region of interest from the background to obtain a final fine segmentation image. With the technical solution, the trained coarse detection network is used to infer the input image, quickly locate the position of the aneurysm target, and then the fine segmentation network is used to finely segment the region of interest, thereby improving the accuracy and detail expression ability of aneurysm segmentation.